Early detection of VoIP network flows based on sub-flow statistical characteristics of flows using machine learning techniques

Abstract

Network traffic classification plays an important role in the areas of network security, network monitoring, QoS and traffic engineering. In this paper, we design a network traffic classifier based on the statistical features extracted from network flows. Instead of deriving the statistical characteristics per flow, our model make use of features extracted from the first few seconds of each flows. The first few seconds of each flow is divided into overlapping time-based windows. This approach enables our classifier to classify each flow early. Attribute selection algorithms Chi-Square, CON and CFS are used to obtain the optimal subset of features. We give a comparative analysis of the result on the said approach based on the classification algorithms (Decision tree (C4.5), Naive Bayes, Bayesian Belief Network and SVM). We also present a single class classifier implementation of C4.5 algorithm. The experimental results show that the proposed method can achieve over 99% accuracy for all testing dataset. Using the proposed method, C4.5 algorithm delivers high speed and accuracy. By taking inference from these offline classifiers, we build an online standalone classifier using C/C++. We used the following applications: Skype, Gtalk and Asterisk.